Body Sensor Network

There are many advantages to being able to wirelessly monitor a person's physiological data (e.g., heart rate, blood pressure, respiratory rate) during their day-to-day life. Keeping track of a person's heart rate and blood pressure, for example, might enable signs of a developing health condition to be identified. In some examples, it may be useful to be able to monitor a patient's vital signs after they have been released from hospital following an operation. Patients may be released from hospital sooner after an operation if health care professionals felt confident that the patient's physiological parameters could be reliably and accurately monitored at home with an on-body device. This in turn can help to free-up hospital beds and resources.

Devices that can monitor a user's physiological parameters in day-to-day life often suffer from being bulky, expensive, and require re-charging frequently (e.g., every few days). This is because the process of obtaining and processing body sensor data is energy intensive, computationally intensive, and requires large amounts of storage. However, the body sensors themselves usually have a very limited battery life and have limited storage and computational capabilities. In particular, transmitting data wirelessly consumes a substantial proportion of the energy of the sensors. This challenges both the quality of the communicated information and the processing resources available at the sensor. Furthermore, needing to recharge the sensors frequently will reduce the efficacy of the monitoring device as the user will be frequently taking the sensors off. This increases the likelihood that signs of a critical condition may be missed and can be dangerous for users who need near constant monitoring for their health conditions. Furthermore, the user may feel less free to move around if they know that the sensors cannot last a long time away from charging facilities.

Recent work in the field has recognised the benefit of providing near continuous monitoring of a user's biometric data.

The paper “-” by F. Sanfilippo and K. Y. Pettersen, 2015 11th International Conference on Innovations in Information Technology (IIT), Dubai, 2015, pp. 262-266, discloses a preliminary working prototype of a wearable integrated health monitoring system that provides the user with haptic feedback and allows the user to report an emergency. One application discussed is for monitoring the vitals of workers such as offshore operators. The monitored data is sent wirelessly to cloud-based data storage, however, the energy constraints on the system are not discussed.

Rinicare have developed the STABILITY system which monitors a patient's condition post-surgery. This provides a personalised score from all currently recorded patient data that predicts the likelihood of an improvement or deterioration in the patient's condition. However, this system does not consider the energy efficiency of capturing physiological data.

ADAMM intelligent asthma monitoring is a wireless wearable device that adheres to the skin around the chest and monitors respiratory rate, oxygen levels, pulse, blood pressure and body temperature. It uses machine learning to analyse the user's data to alert users to risks of asthma attacks. The device is intended to be recharged overnight in a cradle next to the user's bed.

There is provided a system for adaptively capturing physiological data using compressive sensing, the system comprising:

The processor may be configured to approximate each physiological signal from the sampled data by, using the respective sensing matrix, solving for a sparse representation of the physiological signal by applying compressive sensing to the respective sampled data.

The processor may be configured to approximate each physiological signal from the sampled data using interpolation or principal component analysis.

The processor may be configured to update each sensing matrix by determining a respective basis that sparsely represents the respective approximated signal and determine a respective updated sensing matrix that is incoherent with the determined basis.

Solving for a sparse representation of each physiological signal may be performed using the respective basis previously determined to sparsely represent the previously approximated physiological signal.

The processor may be configured to infer changes to the respective physiological signal based on a measure of change of the respective basis each time the respective sensing matrix is updated.

For each physiological signal, the respective basis may be determined using a sparse coding algorithm that finds a sparser basis with which to represent the respective approximated signal, wherein the approximated signal is used as training data for the sparse coding algorithm.

The identified correlations may be used as an additional constraint in the sparse coding algorithm.

Compressive sensing may be applied using an l-norm so as to solve for the sparse representation of the physiological signal.

The processor may identify a measure of correlation between two or more physiological signals.

The processor may update one or more sensing matrices in dependence on whether the measure of correlation between two or more physiological signals deviates from an expected value by more than a predefined threshold.

The processor may update one or more sensing matrices in dependence on whether the rate of change of the measure of correlation between two or more physiological signals deviates from an expected value by more than a predefined threshold.

The processor may update one or more sensing matrices to sample less frequently if the measure of correlation between the respective physiological signals exceeds a predefined threshold.

The processor may update one or more sensing matrices to sample more frequently if the measure of correlation between the respective physiological signals falls below a predefined threshold. The measure of correlation may be one or more of: Pearson's correlation coefficient, Spearman's rank correlation coefficient, or a regression line.

One of the sensors may be an accelerometer configured to measure movement of the user.

The processor may be configured to compare correlations between the accelerometer data and the other physiological signals measured by the system so as to correct for artifacts in the physiological signals caused by human movement during sampling of the sampled data.

For each physiological signal, the basis may be determined using the K-SVD machine learning algorithm.

The physiological signals may be sparse signals such that each physiological signal of length N can be represented as a linear combination of K basis vectors, wherein K<<N.

The processor may comprise a receiving device configured to attach to a user. The receiving device may receive sampled data from each of the plurality of sensors.

There is provided a method for adaptively capturing physiological data using compressive sensing, the method comprising:

Approximating each physiological signal from the sampled data may comprise using interpolation or principal component analysis.

Updating each sensing matrix may comprise determining a respective basis that sparsely represents the respective approximated signal and determining a respective updated sensing matrix that is incoherent with the determined basis.

Solving for a sparse representation of each physiological signal may be performed using the respective basis previously determined to sparsely represent the previously approximated physiological signal.

The method may further comprise identifying changes to the respective physiological signal based on a measure of change of the respective basis each time the respective sensing matrix is updated.

Determining a respective basis for each physiological signal may comprise using a sparse coding algorithm that finds a sparser basis with which to represent the respective approximated signal. The approximated signal may be used as training data for the sparse coding algorithm.

The identified correlations may be used as an additional constraint in the sparse coding algorithm.

Identifying one or more correlations may comprise determining a measure of correlation between two or more approximated physiological signals.

One or more of the sensing matrices may be updated in dependence on whether the measure of correlation between two or more physiological signals deviates from an expected value by more than a predefined threshold.

One or more sensing matrices may be updated in dependence on whether the rate of change of the measure of correlation between two or more physiological signals deviates from an expected value by more than a predefined threshold.

One or more sensing matrices may be updated to increase the sampling frequency of one or more respective physiological signals if the measure of correlation between two or more physiological signals exceeds a predefined threshold.

One or more sensing matrices may be updated to decrease the sampling frequency of one or more respective physiological signals if the measure of correlation between two or more physiological signals falls below a predefined threshold.

Any one or more feature of any aspect above may be combined with any other aspect. These have not been written out in full here merely for the sake of brevity. The features of the above system claims may be implemented as equivalent method claims and vice versa.

The inventors of this application have recognised that there is a large amount of redundancy in a user's physiological data when the user is being monitored constantly. A large amount of the data being transmitted from body sensors will convey essentially the same information and so is often not particularly useful. For example, when a user is sleeping, their blood pressure is likely to remain relatively constant throughout the night. If this is the case, then there is no need to take a measurement of the user's blood pressure every minute. Instead, taking a measurement every half an hour would produce the same picture of the user's state.

Reducing the number of measurements taken by each sensor will increase the battery life of the body sensors in two-ways. Firstly, the act of taking the measurement uses battery power. Secondly, and more importantly, the act of wirelessly transmitting the sampled data is very energy intensive. So, it would be advantageous if the sensors could take fewer measurements so that less data is wirelessly transmitted and/or so that the sensors transmit data less frequently. However, there is a risk, when sampling less frequently, that vital information might be missed by the sensors. So, there is a balance to be struck in taking as few measurements as possible whilst not losing important information from the user's physiological signals.

Furthermore, it would be advantageous if the sensors were able to adapt the number of samples taken depending on the state of the user and the user's activity history. For example, some users may have a more erratic heart rate than other users, and may need their heart rate to be monitored more closely than other users. When the users need the most monitoring may be dependent, for example, on the time of day or their physical state (e.g. whether they are exercising or at rest). There are likely to be patterns in a user's day that correspond to changes in their physiological state. In general, it would be advantageous if the sensors were able to automatically adapt the frequency of sampling to best suit the user.

This application uses compressive sensing technology to reduce the energy consumption of the sensors in a wireless body sensor network (WBSN) by reducing the number of samples taken and wirelessly transmitted at each sensor, without compromising the quality of information obtained by the sensors.

shows an example systemcomprising a plurality of sensorsand a processor. The systemis an example of a wireless body sensor network. Each sensorsamples a physiological signal. A physiological signal is produced from measuring physiological data over time. For example, heart rate may be the physiological data being sampled, and the physiological signal is produced from measuring that heart rate over a certain period of time. Physiological data (herein equivalently referred to as physiological signals) may be defined as data relating to the body of the user. The user is likely to be a human but it may be any animal that has measurable physiological data. Examples of physiological data are heart rate, blood pressure, skin temperature, respiratory rate, and limb or body movement.

The plurality of sensorsmay be attached to a userat different positions on the user's body (as shown in) so that different physiological signals of the user can be measured. The sensors are designed to be continuously worn by the user during day-to-day life.

Each sensoris arranged to measure a physiological signal. One or more of the sensors may measure different physiological signals. Some sensors may measure the same physiological signal as other sensors. This may help improve the accuracy of the physiological signal being measured. Some sensors may measure several physiological signals at one time. The specific physiological signal (or signals) sampled by each sensor may depend on where each sensor is positioned on the body and what physiological signal each sensor is designed to measure. For example, a first sensor may measure the user's heart rate—e.g. at a chest strap adapted to measure the electrical activity of the heart. A second sensor may measure the user's blood oxygen levels—e.g. over an artery at the wrist. A third sensor may measure the user's body movements—e.g. at a limb of the user. One or more of the first, second and third sensors could share the same sensor hardware and/or be located at the same point on the user's body—for example, sensor hardware at the wrist could comprise the first sensor for measuring heart rate and the second sensor for measuring blood oxygen.

shows an example sensorcomprising a measurement unit, a transceiver unitand a memory unit. The measurement unitenables the sensor to capture the respective physiological signal. For example, the measurement unitmay be a heart rate monitor or blood oxygen monitor. The measurement unitmay vary between sensors depending on what physiological signal each sensor is adapted to sample. The sensorsmay comprise a transceiver unit. The transceiver unittransmits the sampled data (e.g. to the processor) and receives data (e.g. from the processor). In some examples, the sensorsmay optionally comprise a memory unit. The memory unitmay store sampled data until the data is transmitted to the processor.

The sensors may optionally perform some simple signal processing prior to transmitting the sampled data to the processor. The transmitted sampled data may comprise the raw measurements captured by the sensor and/or values derived from such raw measurements. For example, the sensors may average one or more measurements prior to transmission. In addition to the sensorsshown in, the systemmay further include conventional sensors for capturing physiological data which do not sample according to sampling patterns in the manner described herein.

The data sampled by the plurality of sensors is transmitted to the processor, as signified by the arrow A in. The processor transmits sampling patterns to the sensors, as signified by the arrow B in.

A sampling pattern describes when a measurement (equivalently referred to as a sample) is to be taken by a sensor. Each sensor receives a respective sampling pattern that determines when that sensor is to take samples. For example, a sampling pattern may determine a frequency at which the sensor takes samples—e.g. every 30 seconds. Generally, the sampling pattern may indicate to the respective sensor in any suitable manner when it is to take a measurement. The sampling pattern need not cause the sensor to take a sample at a uniform rate (e.g. every 30 seconds).

A sampling pattern may indicate when measurements are to be taken within some window of time represented by the sampling pattern. For example, a sampling pattern may represent, for a sequence of time slots, whether or not samples are to be taken in each of those time slots. For example, the sampling pattern may comprise a set of randomly populated binary values, each value indicating whether or not a measurement is to be taken in the respective time slot. In some cases, a sampling pattern may determine that the sensor is not to take any samples.